AIT flashcards

32 cards. Read the term and the definition here. The deck at the top of this page is optional.

Rational agent

An agent that selects actions expected to maximize its performance measure given the percept sequence and built-in knowledge. Rationality depends on the performance measure, prior knowledge, and available actions.

PEAS description

A framework specifying an agent task as Performance measure, Environment, Actuators, and Sensors. It forces explicit definition of success criteria and interaction channels before designing the agent.

Admissible heuristic

A heuristic that never overestimates the true cost to reach the goal. With an admissible heuristic, A* search is guaranteed optimal because f(n) never exceeds the true cost through n.

First-order logic

A knowledge representation adding quantifiers and relations over objects to propositional logic. It expresses statements like 'all humans are mortal' but entailment is only semi-decidable.

Strong versus weak AI

Strong AI asserts machines can possess genuine general intelligence or mind; weak or narrow AI asserts machines simulate intelligent behavior without understanding. Most deployed systems are narrow.

Bias-variance tradeoff

Total expected error decomposes into bias squared, variance, and irreducible noise. Simple models yield high bias (underfitting); complex models yield high variance (overfitting).

L1 versus L2 regularization

L1 adds lambda times the sum of absolute weights and drives some weights to zero for feature selection. L2 adds lambda times the sum of squared weights and shrinks weights smoothly.

Precision and recall

Precision equals true positives divided by predicted positives (TP/(TP+FP)). Recall equals true positives divided by actual positives (TP/(TP+FN)). F1 is their harmonic mean.

Stratified k-fold cross-validation

A resampling method splitting data into k folds while preserving class proportions in each fold. It gives a more reliable generalization estimate for imbalanced classification than plain k-fold.

Data leakage

Occurrence when information from the target or test set enters training features, inflating apparent performance. Prevent it by splitting data before fitting scalers, encoders, or imputers.

Supervised versus unsupervised learning

Supervised learning fits a mapping from labeled inputs to targets. Unsupervised learning finds structure such as clusters or latent factors in unlabeled data without target labels.

Backpropagation

An algorithm computing gradients of the loss with respect to each weight by applying the chain rule backward through the network. Gradients then drive weight updates via an optimizer.

ReLU activation

A nonlinearity defined as max(0, x). It is cheap to compute and mitigates vanishing gradients compared with sigmoid or tanh, though it can produce dead units.

Convolutional weight sharing

Convolutional layers reuse the same filter weights across spatial positions, drastically reducing parameter count versus fully connected layers and encoding translation equivariance in images.

LSTM gates

Long short-term memory networks use input, forget, and output gates to control information flow through a cell state, mitigating vanishing gradients in long sequences.

Dropout regularization

A technique randomly zeroing a fraction of activations during training, forcing redundant representations and reducing overfitting. At inference, dropout is disabled and activations are scaled.

Softmax function

A function converting a vector of logits into a probability distribution using exponentiation normalized by the sum of exponentials. It is standard for multiclass classification outputs.

Overfitting detection

Overfitting is indicated when training loss keeps falling while validation loss rises. Remedies include more data, regularization, early stopping, and reducing model complexity.

Transformer self-attention

Self-attention computes softmax(QK^T divided by the square root of d_k) times V, letting each token weigh all others in parallel. This enables long-range dependency modeling without recurrence.

Subword tokenization

Schemes such as byte-pair encoding and WordPiece split text into frequent subword units. They balance vocabulary size against out-of-vocabulary handling for rare or novel words.

Retrieval-augmented generation

A pattern retrieving relevant external documents and inserting them into the prompt before generation. It grounds model outputs in current sources and reduces hallucination without retraining weights.

Hallucination

Fluent model output that is factually unsupported or fabricated. Mitigations include retrieval grounding, citation requirements, and automated or human verification of claims.

Parameter-efficient fine-tuning

Adapting a pretrained model by training small added parameters, such as low-rank adapters (LoRA), instead of all weights. It reduces memory and compute cost while retaining task performance.

Perplexity

An intrinsic language-model metric equal to the exponentiated average negative log-likelihood per token. Lower perplexity indicates the model assigns higher probability to held-out text.

Demographic parity

A fairness criterion requiring equal positive prediction rates across protected groups. It ignores true labels, so it can conflict with equalized odds and individual fairness.

SHAP explanations

A post-hoc method attributing a prediction to features using Shapley values from cooperative game theory. It provides additive, locally faithful attributions but can be computationally costly.

Differential privacy

A guarantee that query or model outputs change little when any single record is added or removed, bounded by parameter epsilon. Smaller epsilon means stronger privacy and more noise.

EU AI Act risk tiers

A regulatory classification of AI systems into unacceptable, high, limited, and minimal risk tiers. Each tier carries obligations ranging from prohibition to transparency requirements.

Model card

A document reporting a model's intended use, training data, performance across subgroups, and known limitations. It supports transparency and informed downstream deployment decisions.

Concept drift

A change over time in the relationship between input features and the target, degrading model performance. It requires monitoring and retraining, unlike pure input distribution shift.

Feature store

A centralized system computing and serving features consistently for both training and inference. It reduces train-serve skew and enables feature reuse across models.

Model quantization

Reducing the numerical precision of model weights and activations, such as float32 to int8. It lowers memory and latency at some cost to accuracy.